What is the AI Agent Task Manager?
It is a free, browser-based task manager for multiple AI coding agents running in parallel — also known as the AI Agent Task Board because it presents the work as a kanban board. Each column is a workstream (one agent session); each card is a task with a title and an instructions body. You can use it as a plain manual board, or connect autonomous agents that pull tasks from it and report progress back via a simple API.
Is it free, and do I need to sign up?
The board is free. You can use it anonymously with no signup — it saves to your browser via IndexedDB. To connect autonomous agents (which need a private API key) and to sync your board across devices, sign in with Google.
How do agents actually connect to the board?
Click "Connect an agent" on any workstream, generate your personal API key, and copy the bootstrap prompt it gives you. Paste that into a fresh Claude Code, Codex, or Cursor session. The prompt tells the agent to call the board over HTTPS (with curl): pull its next task, do it, mark it complete, and repeat — fully autonomously.
What happens when an agent gets stuck and needs me?
This is notify-only by design. The agent flags the task, that column turns amber ("needs you"), and your browser fires a notification so you know which terminal to go to. You answer the agent directly at that terminal; it then clears the flag and resumes itself. The board is your at-a-glance signal, not a chat relay.
Is the API key secure?
The key is a long random token shown only once at generation; only a SHA-256 hash is stored. It is scoped strictly to your own board — it can read and write only your workstreams and tasks. Regenerate it anytime (which instantly invalidates the old one). Treat it like a password: keep it out of public repos and shared logs.
Which agents and tools work with it?
Anything that can make an HTTP request. It is built and tested for Claude Code, OpenAI Codex, and Cursor in full-auto mode, but the protocol is plain REST (pull / attention / resume / complete) so your own scripts or any agent framework can drive it too.
Can I run several sessions on the same project?
Yes. Workstreams are independent columns — make as many as you like for one project (e.g. "frontend", "backend", "tests") or spread them across different projects. Each has its own task queue and its own connected agent.
Is my data private?
Anonymous boards never leave your browser. Signed-in boards are stored in Codersera’s database scoped to your account. We never share or sell your data, and never use it to train models. Task instructions are stored as plain text so your agents (and you) can read them back exactly.
Do I have to automate it? Can I just use it as a manual board?
Absolutely use it manually. Plenty of people just want one place to see what each of their parallel agent sessions is doing and tick tasks off by hand. The agent automation is an optional power-up — the board is useful on its own as external memory across sessions.
What is the difference between this and the Task Tracker?
The Task Tracker is for tracking YOUR own tasks — a personal Kanban with Backlog → Today → Doing → Done, plus a List view and a Matrix view. The AI Agent Task Manager is for tracking AGENT tasks: each column is one autonomous agent session, and the cards are work the agent picks up and runs. Different mental model, different surface. Many people use both side-by-side — your tasks on /tools/todo-tracker, your agents here.
How does this pair with the Impact-Effort Matrix?
The Impact-Effort Matrix (/tools/impact-effort-matrix) is where you decide WHAT is worth queueing for an agent. Drag your candidate tasks onto a 2×2 of impact and effort; Quick Wins and Major Projects come out the right side; Fill-Ins and Time Wasters get cut. Then move the survivors onto the agent board as actual workstream cards. Triage on the matrix, execute here.
Can I drag and reorder tasks?
Yes — every TODO card is draggable. Drop it above or below another card to reorder within a column, or drop it on a different column to move it across workstreams in one motion. Running and "needs you" tasks are intentionally locked so an in-flight agent can not be yanked sideways. The drop indicator (a blue line) shows exactly where the card will land.
Can I see and re-copy my agent API key later?
Yes. Open Connect an agent → your key sits in a Reveal / Copy box. The bootstrap prompt below also auto-embeds the persisted token on every visit, so you do not have to keep the original copy-once value around. Regenerate any time to rotate (which instantly invalidates the old key).
Can two agents share the same workstream?
Technically yes — the protocol does not enforce one-agent-per-stream — but it is not the design. The top-of-queue claim is atomic, so two agents pulling at the same time will each get a different task with no double-work. That said, you lose the mental model the board is built around (one column = one terminal). If you have two agents, give them two columns.
What happens if I close the browser tab? Does the agent keep working?
Yes. The agent runs in its own terminal session on YOUR machine, not in the browser. The browser tab is only a status surface — close it and the agent keeps pulling, doing, and completing tasks autonomously. Reopen the tab any time to see the up-to-date state. With Web Push enabled, even a closed tab can still ping your OS when an agent flags attention.
How does long-polling work — do agents busy-poll the server?
No. The /api/workstreams/agent/next endpoint accepts a wait=<sec> parameter (clamped to 30 minutes). The connection stays open up to that long; the moment a new task lands on your stream, an in-process pub/sub wakes the parked connection and returns the task immediately. The bootstrap prompt uses wait=1800, so an idle agent costs one HTTP connection, not constant polling.
What is the difference between the "result" field and the instructions body?
Instructions = what YOU write for the agent, the input. Result = what the AGENT writes back when it marks the task complete, the output. The result shows as a one-line summary on the done card so you can scan a column of finished work without opening anything; click the ⛶ icon for the fullscreen viewer with both fields.
Can I use it on mobile?
Yes — the board, status chips, drag-and-drop, and notifications panel all work on mobile browsers. The drag-and-drop is HTML5-native, so it works best with stylus / mouse and is fiddly with touch (use the task edit modal to move cards across columns on touch devices). Push notifications need a modern mobile browser with Web Push support (Chrome on Android, Safari on iOS 16.4+).
Is there a free tier limit?
No artificial caps on the free tier. You can run up to 24 workstreams concurrently per board (a soft cap to keep the UI usable), unlimited tasks per stream, and unlimited tasks completed lifetime. The only billed components in the system are the Web Push delivery (free via VAPID) and the long-poll connections (free, served from the same Next.js process as the page itself).
Does it work with self-hosted Claude / open-source agents?
Yes. The board only knows HTTP — anything that can make a curl request can drive it. Self-hosted Claude via the Anthropic API, OpenAI-compatible local models via vLLM, Ollama agents, custom Python scripts wrapping a local LLM — all work. The bootstrap prompt is a guide; the protocol is the contract.
